DeepSeek’s Modular Approach: A Wake-Up Call for AI Incumbents

The AI industry has long been dominated by the belief that cutting-edge hardware is essential for building powerful models. OpenAI, Anthropic, and other major players have invested billions in the latest Nvidia chips, justifying their high costs with the argument that superior performance requires top-tier infrastructure. However, DeepSeek has shattered this assumption by demonstrating that advanced AI can be developed using older Nvidia chips. This revelation has triggered an aha moment—one that incumbents are struggling to accept.

The Reality Check: DeepSeek Has Done It

DeepSeek has provided clear evidence that it is possible to build competitive AI models without relying on the newest and most expensive hardware. While its competitors continue to spend vast sums on high-end GPUs, DeepSeek has successfully trained its model using older Nvidia H800 chips. This approach is not just cost-effective; it is also remarkably efficient.

There are at least 11 different ways to build AI infrastructure, and DeepSeek’s method is proving to be one of the most effective at this stage of technological development. By adopting a modular approach, DeepSeek has managed to optimise computational resources while maintaining high performance. Instead of treating AI infrastructure as a monolithic system that requires constant hardware upgrades, DeepSeek has broken it down into interchangeable modules that work together seamlessly.

Understanding the Modular Method

To put it simply, DeepSeek’s modular system is like building with LEGO blocks rather than sculpting a single, unchangeable structure. Traditional AI models are often designed in a way that ties them heavily to specific hardware, making them expensive to scale and upgrade. DeepSeek, on the other hand, has developed an approach where different parts of the system can be improved or replaced without overhauling the entire infrastructure. This flexibility allows for better efficiency, lower costs, and greater adaptability—all without sacrificing performance.

By using a Mixture-of-Experts (MoE) model, DeepSeek ensures that only the necessary components of the AI system are engaged at any given time, rather than activating the entire network. This significantly reduces the computational burden, allowing DeepSeek to achieve comparable, if not superior, results to its competitors—without the need for the latest GPUs.

Why Incumbents Must Take Note

The biggest mistake OpenAI and other high-priced incumbents can make now is to ignore DeepSeek’s breakthrough. If they continue to insist that expensive hardware is the only way forward, they risk becoming obsolete. DeepSeek’s success proves that cost-efficiency and smart architecture matter just as much—if not more—than raw computational power.

More importantly, DeepSeek has embraced open-source development, further disrupting the industry. By making its technology accessible, it is fostering a new wave of innovation that could accelerate AI progress beyond what closed systems can achieve. If OpenAI and its peers do not find a way to differentiate themselves—whether through unique applications, better user experience, or enhanced adaptability—they could quickly lose their competitive edge.

The Future of AI Infrastructure

DeepSeek’s modular approach is a game-changer. It challenges the narrative that only the wealthiest AI firms can build state-of-the-art models, proving that strategic efficiency can rival brute-force computing power. As AI continues to evolve, companies that cling to outdated assumptions will find themselves falling behind.

The industry must now ask itself a critical question: Will it adapt to this new paradigm, or will it remain in self-denial while more agile players like DeepSeek reshape the landscape? One thing is certain—AI infrastructure is no longer just about having the most powerful chips. It is about building smarter, more efficient systems that can evolve with technological progress.